[Perf] Ensure async h2d copies are pinned in more places (#55202)
This commit is contained in:
@@ -190,6 +190,7 @@ def test_fast_plan_decode_warmup_uses_full_plan(dtype: torch.dtype) -> None:
|
||||
indptr_cpu=kv_indptr,
|
||||
indices=kv_indices,
|
||||
last_page_len_cpu=kv_last_page_lens,
|
||||
seq_lens_cpu=torch.tensor(kv_lens, dtype=torch.int32, device="cpu"),
|
||||
num_qo_heads=num_query_heads,
|
||||
num_kv_heads=num_kv_heads,
|
||||
head_dim=head_size,
|
||||
@@ -268,6 +269,7 @@ def test_fast_plan_decode_matches_full_plan(
|
||||
indptr_cpu=kv_indptr,
|
||||
indices=kv_indices_buf,
|
||||
last_page_len_cpu=kv_last_page_lens,
|
||||
seq_lens_cpu=torch.tensor(kv_lens, dtype=torch.int32, device="cpu"),
|
||||
num_qo_heads=num_query_heads,
|
||||
num_kv_heads=num_kv_heads,
|
||||
head_dim=head_size,
|
||||
|
||||
Reference in New Issue
Block a user